CRFLOE: Context Region Filter and Relation Word Aware for Document-Level Relation Extraction

The goal of Document-level Relation Extraction (DocRE) is to identify all entity pair relations from a document in one pass. One challenge faced by DocRE is mining the decisive context of entity pair relations from the document. Nevertheless, it is noteworthy that even among human-annotated key sentences, only a limited subset can effectively function as pivotal contextual information. To enhance the acquisition and utilization of crucial contextual information, we propose a pioneering framework named CRFLOE (Context Region Filter and Relation Word Aware). This framework operates by segmenting the context with a focus on entities, subsequently computing region-level attention scores, and filtering out contextual information with low scores at a coarse-grained level. Following this filtration, fine-grained score adjustment is conducted on the remaining candidate regions. Furthermore, we introduce additional prior knowledge of relation word to calibrate the final relation scores, further improving the performance of our model. We conducted experiments on four DocRE datasets: DocRED and ReDocRED, which are two recently released large-scale datasets, as well as GDA and CDR in the biomedical field. Our findings demonstrate the effectiveness of CRFLOE.